Imagine one tenth of a liquid drop dispersed through two Olympic-sized swimming pools. Taking a typical drop as about 0.05 millilitres and the pools as roughly 2.5 million litres each, that dilution is close to one part per trillion by volume. For some odorants, dogs have detected concentrations approaching that scale under controlled conditions.
Researchers have used that sensitivity to investigate an idea that once sounded improbable: disease may change the chemical mixture released in a person’s breath, and a trained dog may be able to recognize the difference. Controlled experiments have reported dogs distinguishing breath from people with lung cancer and colorectal cancer from control samples.
The evidence is striking, but the boundary matters. These were scent-detection studies, not dogs independently diagnosing patients in ordinary clinics. They demonstrate a detectable biological signal and a possible route toward new breath tests. They do not replace CT screening, colonoscopy, pathology or medical assessment.
One part per trillion is not one universal limit
A review of canine detection and disease-associated volatiles cites a demonstrated lower limit near one part per trillion and translates it as about one drop in 20 Olympic pools. One tenth of a drop in two pools is the same ratio. It is a useful picture of scale, but it needs qualifications.
There is no single canine detection threshold. It changes with the chemical, whether it is presented in air or liquid, the equipment, the dog, training and the rule used to define a successful detection. “Approaching” one part per trillion is therefore more accurate than saying every dog can smell every substance at that concentration.
Laboratory thresholds also measure a simpler question than cancer detection. A dog may be asked whether a known odorant is present against a clean background. Human breath contains hundreds of volatile compounds, moisture and chemical variation from diet, medications, smoking, infection and the surrounding environment.
A nose built to separate breathing from smelling
Dogs do more than carry extra smell receptors. Their nasal airflow is organized around repeated sampling. A fluid-dynamics study of canine olfaction describes how inspired air divides into a respiratory stream and an olfactory stream directed toward sensory tissue in a recessed part of the nose. During sniffing, exhaled air leaves through side slits, helping pull fresh odor toward the nostrils.
Behind that airflow sits a large, folded olfactory epithelium containing hundreds of millions of sensory neurons. The brain then interprets activity across receptor combinations. This allows a trained dog to respond to a chemical pattern rather than requiring researchers to identify one decisive “cancer molecule” in advance.
Training is essential. Dogs learn that one class of sample earns a specific alert, such as sitting or lying down, while control samples should be ignored. Reliable experiments must ensure that neither the handler nor the observer can unconsciously reveal which sample is positive.
Breath samples from people with lung cancer
One influential 2006 experiment trained five ordinary household dogs using food rewards and a clicker. They learned to distinguish exhaled breath from 55 people with lung cancer and 31 with breast cancer from samples supplied by 83 healthy controls. Testing used samples the dogs had not encountered during training, while handlers and observers were blinded.
For lung cancer, the reported sensitivity and specificity were both about 99 per cent. Sensitivity describes the share of cancer samples correctly flagged; specificity describes the share of controls correctly rejected. The result established an important proof of concept but came from a deliberately structured study, not a screening clinic.
A separate 2017 lung-cancer experiment exposed one trained dog to exhaled-gas samples from 85 patients with lung cancer and 28 controls. Across repeated presentations, the researchers reported 95 per cent sensitivity and 98 per cent specificity.
Results are not uniformly that high. In a newer study involving seven dogs and 154 patients attending a general respiratory clinic, mean detection of cancer-positive breath samples was 78 per cent, while mean correct classification of non-target breath samples was 68 per cent. That more screening-like population illustrates why performance estimates depend heavily on who is tested and how controls are chosen.
The colorectal cancer result came from one Labrador
Colorectal cancer supplied another memorable breath result. A Japanese study collected exhaled breath and watery stool before participants underwent colonoscopy. During each breath trial, a specially trained Labrador first smelled a standard colorectal-cancer sample, then searched five boxes containing one cancer sample and four controls.
Across 33 breath groups, the dog achieved 91 per cent sensitivity and 99 per cent specificity relative to the participants’ conventional diagnoses. Performance remained high for early cancer, and the researchers did not find obvious confounding by smoking or benign colorectal and inflammatory disease in their sample.
Still, the experimental unit was one dog performing a forced-choice task. The animal always searched a lineup constructed to contain a positive sample. Real screening must also handle a stream of people in which almost everyone may be cancer-free, the disease prevalence varies, and no target is guaranteed to be present.
Why a compelling signal is not yet a clinical test
Medical scent studies are unusually vulnerable to hidden cues. Dogs can learn differences in hospitals, storage containers, collection routines or demographic groups instead of cancer chemistry. Small studies can overestimate accuracy, and repeated measurements from the same people are not equivalent to testing the same number of independent patients.
Good designs therefore use samples from previously unseen participants, randomize positions and keep handlers and scorers blind. They also include people with non-cancerous respiratory or gastrointestinal disease, because the clinically useful question is not merely whether a dog can separate advanced cancer from unusually healthy volunteers.
Positive and negative predictive values also depend on prevalence. Even a test with impressive sensitivity and specificity can generate many false alarms when used for a rare disease in a low-risk population. Standard deployment would require reproducible training, certification, rest schedules, blinded quality-control samples and a plan for what happens after an alert.
This is why the older ScienceBlog report on the 2006 breath study is best read as the beginning of a research program, not the arrival of canine oncology. Later work confirms that a signal exists while showing how much results vary across study designs.
The dog’s best role may be teaching machines what to find
The dogs may ultimately function as biological detectors that point researchers toward better instruments. Cancer metabolism, inflammation and interactions with the microbiome can alter volatile organic compounds circulating in blood and leaving through breath. The canine alert says the combined pattern is distinguishable even when scientists do not yet know which molecules carry the most information.
Gas chromatography, mass spectrometry and electronic sensor arrays can then search for those patterns. A machine would be easier to calibrate, reproduce and distribute than teams of working dogs, provided it can match their ability to recognize a complex scent against the noisy background of real human breath.
The scientific achievement is already substantial. Trained dogs have shown that breath can carry enough information to separate some cancer samples from controls, at times with remarkable accuracy. The clinical challenge is harder: identify the signal, prevent shortcuts and confounding, and prove in large prospective populations that the result changes patient care for the better.
Until that happens, a canine alert belongs in research, not in place of a physician’s diagnosis.